DNAS-Net:一种用于肺炎分割的新型密集嵌套网状编码器  

DNAS-Net:A novel dense nested anastomosing encoder for pneumonia segmentation

作　　者：刘庭江 周凯 章毅[1] 徐修远[1] LIU Ting-Jiang;ZHOU Kai;ZHANG Yi;XU Xiu-Yuan(Department of Computer Science,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学计算机学院,成都610065

出　　处：《四川大学学报（自然科学版）》2024年第6期49-58,共10页Journal of Sichuan University(Natural Science Edition)

基　　金：国家自然科学基金(62106163);四川省科技厅重大科技项目(2020YFG0473);四川省自然科学基金项目(23ZDYF0623)。

摘　　要：自2019年新型冠状病毒肺炎迅速传播以来,肺炎的检测和治疗逐渐成为一个热门话题.基于深度学习的计算机辅助筛查作为提高肺炎筛查和临床诊断准确性的辅助手段,受到了广泛关注.然而,由于医学图像中病变的形状、大小和位置存在显著差异,传统的深度学习方法在有限的医学图像数据集上表现不佳.我们提出了一种新的密集嵌套网状肺炎分割模型(Dense Nested Anastomosing Segmentation Network,DNAS-Net),它是一种从CT图像中分割肺炎区域的有效模型.首先,该模型使用注意层次空间金字塔模块(Attentive Hierarchical Spatial Pyramid Module,AHSP)和注意可分离特征金字塔模块(Attentive Separable Feature Pyramid Module,ASFP)建立了一种用于深度特征提取的密集嵌套网状编码器.其次,在跳跃连接中,模型不再只连接来自对应阶段的编码器和解码器,而是使用密集跳跃特征融合模块(Dense Skip Feature Fusion Module,DSFF)来弥合低级和高级特征之间的语义差距,以促进语义分割.大量实验表明,提出的DNAS-Net具有更好的分割准确性.Since the emergence of the COVID-19 coronavirus in 2019,there has been a growing interest in the detection and treatment of pneumonia.Computer-aided screening utilizing deep learning has emerged as a promising tool to enhance the accuracy of pneumonia screening and clinical diagnosis.However,traditional deep learning methods have shown limitations in effectively analyzing medical image datasets due to signifi⁃cant differences in the shape,size,and location of lesions in medical images.We propose a Dense Nested Anastomosing Pneumonia Segmentation model(DNAS-Net),as a novel and efficient method for segment⁃ing pneumonia regions in CT images.The model incorporates a dense nested anastomosing encoder that uti⁃lizes two hierarchical pyramid modules,namely the Attentive Hierarchical Spatial Pyramid module(AHSP)and the Attentive Separable Feature Pyramid module(ASFP),for efficient deep feature extraction.Skip connections are utilized to connect encoders and decoders from corresponding stages and a Dense Skip Feature Fusion module(DSFF)is employed to bridge the semantic gap between low-level and high-level fea⁃tures to enhance semantic segmentation.Extensive experimental results demonstrate that the proposed DNAS-Net achieves higher segmentation accuracy.

关 键 词：深度学习 图像分割 肺炎 多尺度 跳跃连接 

分 类 号：O391.41[理学—工程力学]